Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
A new pragmatic approach to solve the problems of vector
optimization with uncertain parameters
V I Reizlin and A A Nefedova Department of Informatics and System Design, Tomsk Polytechnic University, 30,
Lenina ave., Tomsk, 634050, Russia
E-mail: [email protected]
Abstract. In this paper, we consider the method for solving problems of multi-criteria
optimization with mathematical models that contain a lot of variables. The values of these
variables are not regulated by a decision-maker. We introduce a new concept of ‘tolerance of
the decision variant’. This concept is similar to such concepts as stability, survivability, etc.
1. Introduction
Any decision-making situation is characterized by the following common elements:
a) The set of variables, the values of which are chosen by a decision-maker (hereinafter – DM).
Such variables are called variants of a solution or easy variants.
b) The set of variables, which are not regulated by DM. We call such variables the conditions.
c) The method evaluating quality solutions for each of the conditions. Usually it is one or more
functions depending on the options and conditions.
Now let us turn to the mathematical formulation of the problem.
This paper assumes that the solutions are described by n-dimensional vectors 1,...,
nnx x x R ;
conditions are also vectors 1,...,
mnp p p R . The components of vectors x are design
parameters, and the components of vectors p – external parameters. A pair of vectors ,x p is
called a situation. In other words, the situation is a variant of solution x under condition p.
We also assume that the parametric restrictions are as follows:
, 1, , ,i i ix x x i n
, 1, , .i i ip p p i m
These restrictions define rectangular block B X P in the parameter space n mR R , which we call
an original block. Block X is the block of variants and block P is a block of parameters. In addition to
parametric constraint we shall consider functional limitations 0 ( , ) , 1i ig x p g i r .
In the original block they cut some curved region G, which is called a range of possible situations.
A projection of G to the space of the design parameters is denoted by G|X. The set of G|X is a set of
alternative solutions allowing one to make a choice. Finally, the quality of situation
‘Variant + conditions’ will be evaluated using objective functions (criteria) ( , ), 1if x p i s , which
are defined in B. In other words, under other equal conditions the situation is better if there are more
target functions.
MEACS2015 IOP PublishingIOP Conf. Series: Materials Science and Engineering 124 (2016) 012090 doi:10.1088/1757-899X/124/1/012090
Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distributionof this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
The basic concept for multi-criteria problems is Pareto optimality. Situation ,x p is Pareto
optimal if there is no other situation * *,x p , such that * *( , ) ( , )i if x p f x p for all i and at least for
one of the objective functions this inequality is strict. We need to find a variant of the solution, which
generates an optimal situation for a wide variety of conditions
So, we have the following problem. In area G, we need to select ‘best’ vector x that defines a choice
of the variant, the quality of which is in any way consistent with a set of objective functions. We have
to understand what a ‘best’ vector is. Let us consider first the simplified version of the problem when
external parameters are missing.
2. The existing approaches
The existing methods for solving the multi-criteria optimization can be divided into automatic and
interactive.
2.1. Automatic methods
These methods provide the obtaining of a single compromise point (the point of the Pareto set), which
is considered to be optimal. To obtain this point in automatic methods of the multi-criteria
optimization, the problem is converted into a task with one criterion of efficiency. The most common
methods of this type are the methods of the main criterion, the generalized criterion and goal
programming [1-4].
2.1.1. A method of the main (primary) criteria. The original multi-criteria optimization problem is
reduced to the task of optimizing one of criteria ( )kf x , which is considered to be the most important
one provided that the values of other criteria should be not less than established valuesif .
2.1.2. Methods of a generalized criterion and goal programming. These methods consist in the
convolution of all criteria ( )if x in a single function, which is called a generalized criterion [5].
In these methods, DM chooses not a particular solution but the generalized criterion. Based on this
criterion, the optimal solution will be selected. The obtained solution is considered to be optimal in the
sense of the selected generalized criterion. Here DM separates oneself from the immediate values of
criteria and can only deal with their normalized values using weighting factors that determine the
importance of the criterion. Automatic methods are good because they give one unique solution.
However, these methods have significant drawbacks. Consequently for the method of the main criteria
it is necessary to:
identify a set of the most important criteria, which is not possible in all practical tasks;
set limits for all criteria except for the principal; setting these limitations is difficult for DM if
they are objectively unrelated to the problem statement. In addition, when DM chooses large boundary
values of the criteria, it often provides an ideal point, i.e. a non-existent one. On the other hand, small
values for the constraints resulted from the multi-objective optimization problem lead to a trivial
problem of maximization of the criterion, which contradicts to the formulation of the problem of the
multi-objective optimization.
2.1.3. An illusion of automatic methods.
The methods of the generalized criterion and goal programming create an illusion of obtaining a single
optimal solution, even though DM does not operate directly on the values of functions, but only
normalized ones. It is not possible to prove the correctness of the choice of the generalized criterion.
The same concerns the justification of the reason why the obtained solution better than any other, apart
from the generalized criterion. Automatic methods emerged from disbelief in DM ability to choose the
desired compromise. Choosing the main or generalized criterion, generally speaking, is arbitrary.
Disadvantages of automatic methods have increasingly convinced us of the need for human
MEACS2015 IOP PublishingIOP Conf. Series: Materials Science and Engineering 124 (2016) 012090 doi:10.1088/1757-899X/124/1/012090
2
involvement in decision-making. Therefore, in recent times interactive methods have been widely
adopted.
2.2. Interactive methods
The essence of interactive methods is the same: DM implements a directed search of compromise
points to select the point that best meets DM preferences. For compromise points at each step of the
procedure automatic methods have been used. The procedure is over when one of the points will
satisfy DM requirements. This point is considered to be optimal. However, the analysis of existing
interactive methods showed the presence of a paradox associated with the assumptions about DM
knowledge: on the one hand, DM decides the best compromising point, i.e. she or he has sufficient
knowledge to select it, and on the other hand, it is assumed that DM is not able to clearly indicate what
solution she or he wishes to obtain [6].
In [6, 7] there are methods proposed for solving the multi-criteria optimization that require DM to
set the desired values of the objective functions. If the point of compromise satisfying specified
requirements is absent, the methods require DM to specify setpoints criteria. On the basis of such
specification a compromising region narrows until it turns into a point. The proposed procedures are
quite flexible, but they demand a very large number of one-criterion extreme tasks for execution of the
decision and, therefore, they are only effective for problems with a small number of criteria.
In [8] an interactive method based on sensing of the parameter space was proposed, that is,
replacing of the original continuous area of its discrete representation. Moreover, in this method it is
assumed that DM has an active dialog with PC. This method has significant advantages, which can be
extended to the solution of problems with external parameters.
Let us consider the basic techniques used in the practice of uncertainty removal in case of problems
with external parameters. In fact, there are only two.
2.3. A minimax approach
Selection is considered to be optimal if it is optimal for the worst case. In other words, external
parameters are excluded from the task by replacing objective function ,if x p with
criterion *( ) min ( , )ip
f x f x p and the optimization problem is solved in terms of certainty. The
advantage of this approach is that the obtained result is guaranteed, but only in case a solution exists.
This approach is very inflexible; it requires a large amount of computation and, besides, it is hardly
reasonable to plan the extreme.
2.4. A statistical approach
Selection is optimal if it is optimal average. The basis of this approach rests upon the idea of external
parameters as random variables with certain probability characteristics. The uncertainty is removed if
we replace the objective functions with their mathematical expectations, calculated using the known
distributions of external parameters. However, for the implementation of the statistical approach it is
necessary to construct a probability distribution of external parameters, which is often impossible to
do.
3. A pragmatic approach
In this paper, we propose a pragmatic approach. Multi-objective optimization tasks are solved not as
purely mathematical constructions, but to satisfy very specific needs of design, planning, etc. In order
to achieve practical results it is often enough to find an alternative solution, which perhaps not of the
best, but of a quite acceptable quality, and keep it within the widest possible set of external factors.
Therefore, in each particular situation, the following construction are really meaningful.
MEACS2015 IOP PublishingIOP Conf. Series: Materials Science and Engineering 124 (2016) 012090 doi:10.1088/1757-899X/124/1/012090
3
For each objective function ,if x p we will assign thresholds – ‘worst’ quality indicators which
are acceptable for DM. In other words, we assign a target or criteria constraints. In this case situation
,x p will be suitable if one all inequalities , )( iif x p f is performed.
Set ( ), : , iiF x p G f x fp will be called an effective region or a region of efficiency.
Let us introduce a new concept, somewhat related to such concepts as durability, stability,
insensitivity, etc. Due to the fact that these terms are already engaged and widely used, we will use the
word ‘tolerance’ using the colors for its values, which mean ‘patience’, ‘ability to accept’,
‘indifference’.
Let M be a set of non-zero volume, which lies in starting block B X P , x is a variant of solution.
Let us call set :( , ,)P x M p P x p M as a set of tolerance for variant x on set M.
The ratio of the volume of set ,P x M to the volume of block P (
(, )
, )P x M
T x MP
we will call
tolerance of variant x on set M. The tolerance characterizes the ability of the variant to keep a situation
within set M under control if the conditions will be changed. In other words, it is an indicator of the
stability of options for changing of the conditions.
When keeping our problem in mind, we can assume that tolerance ,T x F is a characteristic of the
ability of variant x to retain the values of objective functions inside initial region G within the limits
specified by vector1 , , sf f f . An option with the greatest tolerance is a pragmatic solution to
the optimization problem.
Obviously, the variant of the decision with tolerance of 1 is approximately optimal for the minimax
approach. From a statistical point of view, tolerance can be interpreted as a probability of options to
keep the situation within the available area. Thus, the concept of tolerance provides the ability to
generalize the minimax and stochastic approach to solving optimization problems with uncertain
parameters.
Let us note that as set F is a part of the field of G, then tolerance ,T x F is a function not only of
vector x, but also of vectors and gf , which define criteria and functional limitations, respectively.
Let functions ,if x p and ,kg x p be continuous in the domain. Then function , , x gfT :
a) is piecewise continuous in variable x,
b) decreases monotonically and is piecewise continuous in variable f ,
c) is non-decreasing and piecewise continuous in variable g ,
d) if all functions ,if x p and ,kg x p are not equal to a constant on the set of non-zero volume,
, , x gfT is a continuous function.
Unfortunately, the calculation of tolerance for different variants of solutions using analytical
methods is possible only for areas with a simple geometric structure, defined by algebraic equations of
the first or second order. But we also need to find an option with the greatest tolerance. From the formal
point of view, we have two different problems: the problem of calculating the measure of sets and the
task of finding the optimum. We can try to solve both these problems, using good enough evenly
distributed sequences [9-11].
We assume that we know how to build the initial portions of uniformly distributed
sequences k kX и P in blocks X and P.
MEACS2015 IOP PublishingIOP Conf. Series: Materials Science and Engineering 124 (2016) 012090 doi:10.1088/1757-899X/124/1/012090
4
Let A be a perfect (coincides with its closure) subset of block P. Let us suppose also that
1 2, ,. . ., NP P P are the starting points of sequence kP , and NА is a set of those that were in A. Then we
have limN
N
AA P
N . For sufficiently large values of N, we have formula
NAA P
N . We will use
this approximation to calculate tolerance.
4. Schemes for problem solving
There are three fundamentally different cases of the general problem formulation of the multi-
objective optimization:
1. Neither the objective functions, no restrictions are independent of the conditions.
2. A part of the objective function depends on the conditions and the other part does not.
3. All objective functions depend on the conditions.
Let us consider each option in detail and offer a scheme of the solution of the corresponding
problem.
4.1. Case 1
This is a classical problem of optimization in the conditions of certainty. We offer methods of its
solution proposed in [7].
In set G, which coincides with the area of possible options, we select the N probe points 1, , NG G
uniformly distributed in the X sequence. At each sampling point values for all criteria 1 , , Nf f are
calculated. We prepare a table of tests for each criterion. In this table the values are arranged in an
ascending order of quality, i.e. the worst values are located before the best values. For large N the edge
values of the test table for each piecewise-continuous criterion are close to its minimum and maximum.
The table of tests allows estimating not only the extreme values of the criteria, but also makes it
possible to judge the frequency of certain values.
Looking through the table of tests, DM assigns threshold criteria, i.e. specifies vector f . Once the
thresholds are assigned, there is a selection of effective points – points satisfying criteria constraints. It
may happen that for given N there is not a single effective point. This means that if a set of effective
options is not empty, then its volume has the order of /G N . In this situation, we can proceed in two
ways: either to change vector f easing any restrictions, or, if the restrictions do not change, to increase
the number of sampling points. If in case of the repeated increase in the number of N the set of efficient
options is still empty, then there is reason to believe that the restrictions are not compatible. Of course,
one cannot exclude that there is an effective point, but if this is the case, then its neighborhood, in
which all the restrictions are applied, has a very small volume. In this case a variant of the solution
corresponding to that point will be unstable, i.e. small violations of the allowances will lead to
efficiency loss.
Once having found a set of viable options, DM can be pleased with the progress considering that any
alternative solution of the problem has been found. However, DM can go further by choosing a built set
of points of Pareto optimal.
4.2. Case 2
This is the most general case of multi-objective optimization problems. A vector of objective functions
consists of two parts: ( ) ( ) ( ), , ,f x p x x p . In blocks X and P we construct a set of sampling
points. We form tables of tests for all f and g functions counting them at the points of set N KX P . Let
us assign a criteria and functional limitations. For criteria that do not depend on the conditions we find
MEACS2015 IOP PublishingIOP Conf. Series: Materials Science and Engineering 124 (2016) 012090 doi:10.1088/1757-899X/124/1/012090
5
effective options in order to choose Pareto optimal from them, and namely for them we can calculate
tolerance. An option with the greatest tolerance will be a solution.
4.3. Case 3
This is another extreme case, in some sense opposite to the first one. Here, all objective functions
depend on the conditions.
In blocks X and P sets and N KX P of sampling points are constructed. We form tables of tests for
all f and g functions calculating them at the points of set N KX P . Then we assign criteria and
functional limitations. For each variant of the area of effective options we can calculate tolerance. A
variant with the greatest tolerance will be a solution.
5. Conclusion
So, in this paper we have presented a method of solution of the multi-objective optimization with
mathematical models that contain many variables, which are not regulated by the decision-maker. This
technique is based on the newly introduced concept of tolerance of variant solution. Schemes for
three different types of tasks are presented. These procedures can be applied for rather a wide range of
problems, but for the problems with a large number of criteria a huge amount of computation may be
required. To remove these restrictions, it is possible to apply a technique of parallel computing, which
is presented in [12].
References
[1] Podinovskii V V 1981 The mathematical theory of decision-making in complex situations
(Moscow, Ministry of Defense of the USSR)
[2] Bekey G A and Masri S F 1983 Math. and Computers in Simulation 25(3) 210–213
[3] Dubov J A, Travkin S I and Iakimets V N 1986 Multi-criteria models for the formation and
selection of system variants (Moscow, Nauka)
[4] Peterson M 2009 An Introduction to Decision Theory (Cambridge University Press)
[5] Mashunin J K 1986 Methods and models of vector optimization (Moscow, Nauka)
[6] Stepanov А V 1991 Math modeling 3(5) 61–73
[7] Sobol I M and Statnikov R B 1981 Selection of optimum parameters in problems with many
criteria (Moscow, Nauka)
[8] Michalevic V S and Volkovich V L 1982 Computational methods for the study and design of
complex systems (Moscow, Nauka)
[9] Kuipers L and Niederreiter H 1974 Uniform distribution of sequences (New York: John Wiley
& Sons)
[10] Orlov V A and Reyzlin V I 2012 Bulletin of Tomsk Polytechnic University 320(2) 24–26
[11] Reizlin V I 2014 Proc. Int. Conf. on Mechan. Engineer., Autom. and Control Systems 6986918
[12] Reyzlin V I and Tartakovsky E A 2012 Proc. of 7th Int. Forum on Strategic Techn. 6357633
MEACS2015 IOP PublishingIOP Conf. Series: Materials Science and Engineering 124 (2016) 012090 doi:10.1088/1757-899X/124/1/012090
6